What Is Jev AI? TypeSafe’s System One Decision Model

Jev AI in one sentence

Jev AI is a non-autoregressive classification model from TypeSafe that reads your input once and returns a typed, calibrated decision — a choice, a score, or a probability — instead of generating text.

That is the whole product. Everything else here follows from that one choice: it is fast because it never generates token by token, cheap because only input is billed, and limited because a model that picks from your options can never explain itself.

How the Jev model differs from an LLM

It does not generate text

A language model produces one token at a time and you parse whatever comes out. The Jev model does not generate: it evaluates your input in a single pass and returns a typed value — one of your options, a number on your scale, or a probability between 0 and 1. No JSON to parse, no schema to validate, no retry loop.

It cannot invent an answer you did not offer

Ask an LLM to sort a ticket into one of four buckets and it can hand you a fifth. Jev cannot. The output type is fixed before the call, so the failure mode changes shape: it will not hallucinate a category, but it will put a ticket in the wrong one without hesitating, and the probability it returns is your only signal. Read that confidence score: a decision model that says it is 51% sure tells you something a language model would have buried in hedging.

The three question types

Every Jev AI request is text plus one or more typed questions, evaluated in a single call. There are three types.

Choice — pick one of your options

You define the options; Jev picks one. The response carries the pick, a probability for every option you listed, and a confidence value for the decision as a whole. Most routing and triage work uses this type: which queue, which department, which of five intents.

Score — place it on your scale

You describe an ordered scale in plain words — low, medium, high — and Jev places the input on it. The score comes back fractional, so 1.81 on a three-point scale means the input sits most of the way toward the top — a plain label would have thrown that away.

Noul — a yes/no probability

Noul is TypeSafe’s name for a single yes/no judgement, returned as a probability from 0 to 1. Is this spam. Does this contain a complaint. Did the customer already cancel. You get back 0.93 rather than the word yes, so you set the threshold instead of inheriting one.

Several questions can travel in one request, evaluated together, for one input charge.

What a System One model is

TypeSafe calls Jev a System One model. The name comes from Daniel Kahneman’s Thinking, Fast and Slow, which splits human thought into System 1 — fast, automatic, intuitive judgement — and System 2, the slow deliberate reasoning you do on paper. A chain-of-thought LLM is the System 2 move: it thinks out loud, at length, and bills you for every word. A System One model does the opposite. It answers immediately, with no reasoning trace, and hands back nothing beyond the answer but a calibrated probability.

One caveat: System One is TypeSafe’s own framing, not an industry term. You will not find it in a standards document or a paper. It is useful shorthand, but say system one outside a TypeSafe context and you will have to explain it. The description that travels is technical — a non-autoregressive classification model with calibrated outputs.

How accurate is Jev AI, really

On TypeSafe’s own four-workflow benchmark, Jev averaged 67.8% agreement with the reference answers. GPT-5.6 Terra scored 67.9%, Opus 5 scored 73.1%, and GPT-5.6 Sol scored 74.1%. On invoice processing specifically Jev managed 61.8% against Sol’s 79.1%. Its best workflow was customer service, at 76.0%.

Two things follow from that. First, Jev is not more accurate than a language model on these tasks — it is faster and cheaper, and on some tasks meaningfully less accurate. Second, the benchmark’s reference answers were produced by averaging the responses of GPT-6 Astra and Claude Fable 5.1. No human labelled anything. So 67.8% measures agreement with two frontier models, not correctness.

That second point matters. A score graded against other models measures consensus, not truth: if both frontier models misread an invoice, an answer that disagreed with them is marked wrong for being right. The honest reading of 67.8% is not “68% accurate” but “agrees with two large models about two-thirds of the time, on a test those models effectively wrote.”

Practically: a green light for routing, a warning for extraction. A 61.8% invoice score means roughly two documents in five come back with a field you would not have picked.

What it costs and how fast it actually is

Jev AI bills $0.042 per million input tokens, with output billed at zero — you pay for what you send, not for what comes back. Against frontier models that becomes a 444× headline; against the small models anyone would actually use for classification it lands nearer 5× to 25× — still real money at volume. The per-call table is on what Jev AI costs.

On speed, TypeSafe’s own material quotes roughly 70 to 500 milliseconds — that number is server-side processing time and does not include network transit. We measured the end-to-end round trip ourselves on 2026-09-20: 20 calls, made from mainland China, routed through OpenRouter (not a direct connection to TypeSafe), using a customer-support-triage payload (one choice question, one score question, one noul question). Fastest call: 309 ms. Median: 423 ms. 90th percentile: 1041 ms. Slowest: 1077 ms. Twenty calls from one location is a small sample, and a connection closer to TypeSafe’s infrastructure would likely be faster — but end-to-end, from a real browser, is the number that actually matters to an integrator, and it is well above the server-side figure.

When to use Jev AI, and when not to

Reach for Jev AI when the decision is small, the answers are known in advance, and you need the verdict now: routing a request to the right queue, moderating content against a fixed policy, scoring a lead or a risk, sanity-checking a large model’s output before it reaches a customer, or classifying a backlog of a million records where per-call price stops being rounding error. In each case the shape of the answer is settled before the call, and a calibrated probability beats a paragraph.

Skip it when you need the reasoning, not the verdict: a decision model cannot tell you why, so anything a human must review, appeal or sign off on wants a model that shows its work. Skip it for anything generative, because the Jev model does not produce text at all. And skip it wherever accuracy is the binding constraint: 61.8% on invoices is the example, and no amount of speed makes up for a field extracted wrong.

Frequently asked questions

What is Jev AI?

TypeSafe’s decision model: it takes your text plus a typed question and returns a typed answer with a probability attached. It does not generate text, and only ever picks from options you supplied.

Is Jev an LLM?

No. Large language models are autoregressive — they generate text one token at a time. Jev evaluates the input in one pass and returns a value of a fixed type.

Can Jev AI hallucinate?

Not in the usual sense: it cannot invent an option you did not define, so it will never return a category that does not exist. It can still be confidently wrong inside the options you gave it — the probability is the warning.

How fast is Jev AI really?

Our own 20-call test on 2026-09-20 — from mainland China, through OpenRouter — came back with a median of 423 ms and a 90th percentile of 1041 ms, end to end. TypeSafe quotes 70 to 500 ms, which is server-side processing only.

How much does Jev AI cost?

$0.042 per million input tokens; output tokens are free. On a 1000-token classification call that is a fraction of a cent, roughly 5× to 25× cheaper than the small language models you would otherwise use.

Is Jev more accurate than GPT or Claude?

No. On TypeSafe’s own benchmark Jev averaged 67.8%, while GPT-5.6 Sol scored 74.1% and Opus 5 scored 73.1%. Jev wins on latency and price, not on accuracy.

What can Jev AI not do?

Write anything, explain its reasoning, call tools, hold a conversation, or handle an open-ended question. If the set of valid answers is not known before the call, this is the wrong model.

Where can I try Jev AI?

In the playground on this site: try Jev AI in your browser— no signup, free runs capped per day. Export the request as curl, Python or TypeScript when it works.